Refreshing Model Pricing

VoiceGateway prices every modality (LLM, STT, and TTS) through voice-prices, a fork of pydantic/genai-prices that covers all three modalities. VoiceGateway no longer keeps any local rate catalogs: rates, source URLs, and verification dates all live in voice-prices.

The wrappers that call into it are:

• src/voicegateway/pricing/llm.py
• src/voicegateway/pricing/stt.py
• src/voicegateway/pricing/tts.py

Each resolves a provider/model id against voice-prices and returns the computed cost. The per-request attribution string is voice-prices@<version> for priced models and voicegateway-local for self-hosted (local/*, ollama/*) models.

When a refresh is required

A rate is refreshed when a provider publishes a price change, or when a model VoiceGateway supports is missing from voice-prices (a pricing call returns None). Freshness is owned by voice-prices: every model entry there carries a prices_checked date and a pricing_source_url, so the verification trail lives upstream rather than in this repo.

How to refresh a rate

1. Confirm the current behaviour from VoiceGateway:

   from voicegateway.inference.pricing import llm, stt, tts
   
   stt.calculate_stt_cost("deepgram/nova-3", 60)   # one minute
   tts.calculate_tts_cost("openai/tts-1", 1000)    # 1000 characters
   llm.calculate_llm_cost("openai/gpt-4o", 1000, 100)

2. Update the rate in voice-prices: edit the model's entry in the relevant provider file under prices/providers/, bump its prices_checked date, and confirm pricing_source_url still points at the provider's price page. Run the voice-prices test suite.
3. Publish a new voice-prices version to PyPI.
4. Bump the pin in VoiceGateway's pyproject.toml (voice-prices>=<new-version>,<0.1) and re-run the pricing tests:

   pytest src/voicegateway/tests/pricing/ -q

Adding a missing model

If a pricing call returns None for a model VoiceGateway should support, the model is not yet in voice-prices. Add it upstream (model id, match pattern, and prices block) following the existing entries for that provider, publish a new voice-prices version, and bump the pin. The coverage tests in src/voicegateway/tests/pricing/test_stt.py and test_tts.py assert that every supported cloud model resolves, so a missing model fails CI until it is added.

Why pricing lives in voice-prices

voice-prices gives LLM, STT, and TTS pricing a single versioned source with a release cadence VoiceGateway can pin against, instead of a hand-maintained catalog that drifts. Refreshing a rate is a voice-prices release plus a pin bump, and the attribution string records exactly which voice-prices version priced each request.